Memory - Computational Explorations in Cognitive Neuroscience - page 312

cognition. Within an intrinsically dynamic neural net-
work framework, these task variations amount to just
that much more data that can be used to test and con-
strain models. Although important for understanding
all aspects of cognition, such models are particularly
important for studying the dynamics of development,
which is all about changes over time.

9.7.1 Recognition Memory

Recognition memory refers to the ability to discriminate
recently experienced stimuli from novel or less recently
experienced stimuli. In the psychological laboratory,
this is typically operationalized by showing participants
a list of words (or pictures), and then testing them af-
ter some delay with some of those “old” words and new
lure words that were not on the previous list. At test,
the participants are asked to say whether a given word
is “old” or “new,” and sometimes they are also asked on
what basis they think the word is “old,” for reasons that
will be clear in a moment.
O'Reilly et al. (1998) have used the hippocampus
model described in this chapter to account for the con-
tribution of the hippocampus to a range of recognition
memory phenomena. The hippocampus can enable the
subject to recollect the experience of having seen an
item before, thus subserving recognition. Thus, if an
input stimulus triggers a strong activation pattern in the
hippocampus that reinstates on the entorhinal cortex
both the input item representation and possibly some
representation of the original study context, the subject
can be reasonably confident in making an “old” judg-
ment.
Indeed, O'Reilly et al. (1998) found that the hip-
pocampus model produced a relatively high-threshold,
high-quality recollective response to test items. The re-
sponse from the hippocampus is “high-threshold” in the
sense that studied items sometimes trigger the recollec-
tion of the probe item, but lures never do so. The re-
sponse is “high-quality” in the sense that, most of the
time, the recollection signal consists of part or all of a
single studied pattern, as opposed to a blend of stud-
ied patterns. The high-threshold, high-quality nature of
recollection can be explained in terms of the conjunc-
tivity of hippocampal representations: Insofar as recol-
lection is a function of whether the features of the test
probe were encountered together at study, lures (which
contain many novel feature conjunctions, even if their
constituent features are familiar) are unlikely to trigger
recollection. Furthermore, the activity patterns that are
recalled are likely to belong together because they were
encoded conjunctively.

9.7 Memory Phenomena and System Interactions

In this final section we will extend the ideas devel-
oped to this point by discussing how a range of dif-
ferent memory phenomena can be explained in terms
of the underlying neural mechanisms explored in this
chapter. Specifically, we will focus on the three spe-
cialized brain areas described earlier: posterior cortex,
frontal cortex (and specifically the prefrontal cortex),
and the hippocampus and related structures. We will
sketch how these areas and their computational proper-
ties could contribute to some well-known memory phe-
nomena, but we will not explore explicit models.
To provide a context for this discussion, it is impor-
tant to understand the nature of the traditional memory
models that have been developed largely from a cog-
nitive, functional perspective (e.g., Hintzman, 1988;
Gillund & Shiffrin, 1984; Anderson, 1983). For the
most part, these models present a monolithic view of
memory where there is a single type of system that con-
tains all of the memory traces, and there is typically
a single canonical representation of a given memory
within the system. In contrast, our framework stipu-
lates that there are multiple interacting memory sys-
tems, each having different characteristics and empha-
sizing different types of information.
This multiple memory systems view has been advo-
cated by a number of researchers (e.g., Schacter, 1987;
Squire, 1987), and is becoming increasingly popular
(Schacter & Tulving, 1994). However, the increased
complexity that comes with this multiple systems view
seems to have thwarted the development of explicit
computational models that instantiate it. Thus, although
still in the early stages of development, the framework
outlined in this chapter can be seen as a first step to-
ward providing a computationally explicit model that
captures the multiple memory systems view.